# Dispatch Task > Determine and dispatch the next authorised action in a governed project. Locates the project's configured task authority, inspects live task, PR, and review state, identifies the responsible role, and either runs the project's configured dispatcher or produces a concrete handoff. Stops when configuration or authority is missing, conflicting, or records that no task is approved. Use only when the user explicitly asks to continue, advance, dispatch, or run the next governed task. - Skill: `tticom/dispatch-task` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add tticom/dispatch-task` - Raw SKILL.md: https://api.skillmd.com/api/skills/tticom/dispatch-task/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: tticom (https://skillmd.com/u/tticom) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/tticom/dispatch-task --- # Dispatch task Find out what the project has authorised next, and route it, without inventing authority. This skill is a generic front door: everything specific to a project (where its task authority lives, which repositories and roles exist, which identities may act, how its runtime dispatches) is read from the **project's own configuration**. Nothing project-specific is built into this skill. **Prerequisite skills:** `identity-safe-git`, `durable-handoff` All AgentOps skills are installed together in one skills directory. If a prerequisite is unavailable, stop and report `REQUIRED_SKILL_MISSING `. This skill is not a task-state system and does not compete with one. The project's runtime owns task state; this skill only reads it and follows what it returns. Do not create queues, task files, status logs, or any state of your own. ## 1. Resolve the project configuration Run the bundled resolver from (or pointed at) the consuming project. It is read-only and never runs anything it reports: ```text python /scripts/dispatch_config.py --project ``` It looks for `agentops-dispatch.json` (or `.agentops/dispatch.json`) at the project root, or a file named by `--config` or `AGENTOPS_DISPATCH_CONFIG`. Read [config-schema.md](references/config-schema.md) for the format. Act on `status`: | Status | Meaning | What to do | |---|---|---| | `READY` | valid configuration with a dispatch command | continue to step 2, then run it in step 4 | | `HANDOFF_ONLY` | valid, but the project has no runtime command | continue to step 2, then hand off in step 4 | | `STOP_NO_CONFIG` | no configuration found | stop and report; ask the project for one | | `STOP_CONFLICT` | more than one differing configuration | stop and report which; never pick one | | `STOP_INVALID_CONFIG` | unreadable or malformed | stop and report the exact error | | `STOP_AUTHORITY_MISSING` | a declared authority document is absent | stop and report the path | | `STOP_INVALID_AUTHORITY` | a declared authority document exists but cannot be read faithfully (not valid UTF-8, or unreadable) | stop and report the path and error; never guess its contents | | `STOP_NO_ACTIVE_TASK` | the task authority says no task is approved | stop and report; do not invent one | Any `STOP_*` status is terminal. Report the status and its message exactly and do nothing further. A recorded backlog item, plan, report, or suggested candidate is never permission to start work. Never reuse an earlier prompt or earlier result: resolve and read state fresh on every invocation. ## 2. Verify identity and role If the configuration names an `identity` profile or role policy, apply `identity-safe-git` before any write: run its identity gate against the profile and keep its role gate for later review or merge operations. An identity mismatch is a no-write stop; never switch accounts or borrow another identity's checkout. The responsible role comes from the project's authority and dispatcher, never from the wording of the request. If the acting identity is not authorised for the role the project assigns, stop and report the mismatch. ## 3. Inspect live state Read, do not assume: - the task authority documents listed in the configuration, at their current revision; - for each configured repository, the open pull requests: number, author, head branch, full head SHA, and state (use the GitHub CLI or the host's API); - for each open PR, the formal reviews and the latest **non-dismissed** review on the **exact live head**, ordered by server timestamp then review ID; a later changes-requested review supersedes an earlier approval on the same head; - author handbacks, kept separate from formal verdicts: a handback comment is never proof that review is pending when a current-head changes-requested verdict exists. A new author head makes any earlier-head verdict historical. If the authority, the PRs, and the reviews disagree about what is current (for example an open PR for a task the authority no longer names), report the conflict and stop. ## 4. Dispatch **`READY`: dispatch through the configured runtime.** Run the resolver's `command.argv` exactly as given, in `command.cwd`, without a shell and without adding or removing arguments. Its output is authoritative: treat its reported state and next action as the answer, and do not reconstruct state in prose or from memory. Only the states the runtime marks as action-authorising permit work, and then only the single bounded action it returns. Any other state is terminal: report it and stop. Repeating the command with unchanged output is idempotent: do not create another task, review, or handback. **`HANDOFF_ONLY`: provide a concrete handoff, and say execution did not occur.** Using `durable-handoff` and [handoff.md](references/handoff.md), state the live facts gathered in step 3, the responsible role, and the exact next action with the commands or steps a person or another runtime should run. Then say plainly: **"No action was executed; this is a handoff only."** Do not perform the action yourself unless the user separately and explicitly asks you to, within the role you are authorised for. ## 5. Boundaries - Permission to dispatch never authorises a merge, a push to a protected branch, a force push, deleting a branch, a credential change, or bypassing a check. - Never merge, self-approve, or promote a candidate into an active task. - Do one bounded action, then stop. Do not chain into the next task. - Delegation to sub-agents is optional: use it only if your runtime provides it and the task authorises it. Do not assume a particular sub-agent tool exists. ## Invocation - **Claude Code:** invoke as `/dispatch-task`, or let Claude select it from its description. - **Codex:** mention it as `$dispatch-task`, or let Codex select it from its description. This skill has no `go` alias. The helper needs only Python 3.11+ and runs natively on Windows, macOS, and Linux; use `python`, or `py -3` where that name is missing.